All Times EDT
Keywords: Biostatistics, Covariate Adjustment, Disease Progression Model, AI, Machine Learning, Synthetic Control
Recruiting patients (pts) into randomized trials costs time and money, creating a need for innovation to improve efficiency of clinical trials. Two methods that have been proposed to achieve this goal involve the use of an external data source. The external data can be used to create a synthetic control arm (SCA) or to build a disease progression model (DPM) to create predicted outcomes. Propensity score matching is a common tool used for analysis for non-randomized trials supplemented by SCA. PROCOVA is a new method, leveraging DPM predicted outcomes as a powerful covariate in an ANCOVA model. With PROCOVA, pts are still randomized to treatment (TX) and placebo, but fewer pts are needed in both arms due to improved power. Data were simulated for an outcome as a function of a measured predictor variable, an unmeasured predictor variable, a TX effect, and random error. We assume 200 pts are prospectively enrolled in a study. In one case, all 200 are treated and an additional 100 pts comprise an SCA. In the other case, the 200 pts were randomized 1:1 and the external data was used to develop DPM predicted outcomes to be used for covariate adjustment. Simulations were conducted varying the size of the TX effect, the amount of bias between the trial data and external data with respect to the unmeasured predictor, and the correlation between the measured and unmeasured predictor. Where the TX effect is zero, PROCOVA consistently preserved type I error, while this is only true for propensity matching if there is zero bias. The problem is exacerbated when correlation between the measured and unmeasured predictor is low but persists even at high correlations less than 1. When the TX effect is positive, both methods improve power. Propensity score matching allows studies to be conducted without randomization, but there is a high price to be paid in loss of type I error control. PROCOVA improves power and maintains type I error control.